Zhang, Ziyang and Angelov, Plamen and Jiang, Richard (2025) Interpretable Deep Learning and its application to Earth Observation. PhD thesis, Lancaster University.
![[thumbnail of 2025ZiyangZhangphd]](https://eprints.lancs.ac.uk/style/images/fileicons/text.png)
Abstract
Deep learning algorithms have revolutionized almost every industry over the last 10 years. They have been used in a wide range of different domains, including image classification and generation, language translation, robotics, health care, and earth observation. However, with the benefit of the promising performance that such deep learning algorithms provides, the models have become larger and more complex. This has led to the current state-of-the-art models becoming ‘back box’ models where people do not understand the reasoning behind a particular decision made by the model. This can be fatal in areas where human life and property are at stake. Floods, as one of the most common natural disasters, cause significant damage to people’s lives and property. Therefore, flood detection has been a key research area in remote sensing. Recently, deep learning based techniques have been extensively used to solve this problem because of their ability to provide high accuracy. However, their black box nature remains controversial and has been criticized by human users. In the case of flood detection, where the safety of a large number of people is at stake, the interpretability and transparency of the algorithm are even more critical. In conclusion, the dissertation aims to provide general interpretable deep learning methods and their application to flood detection. The thesis can be summarized by the following key contributions: - Interpretable method for general image classification. An interpretableby-design algorithm (IDEAL) for image classification has been proposed. IDEAL was shown to provide comparable results compared to standard deep learning architectures without fine-tuning while providing an interpretable decision-making process. - Interpretable methods for semantic segmentation. An interpretable semantic segmentation method (IDSS) for flood mapping has been proposed. It is a prototype-based method that provides both high accuracy and interpretability. Linguistic IF . . . THEN rules together with the prototypes layer, which is the combination of the latent feature space and the raw feature space, have been used to explain the algorithm’s decision-making process. - Interpretable methods for flood detection and flood mapping. An Interpretable Multi-stage Approach to Flood Detection from time series Multispectral Data (IMAFD) has been proposed. IMAFD progressively narrow down the research problem from the time series sequence level to the multi-image level to the image level, providing an automatic, efficient and interpretable approach to flood detection. - Interpretable ensemble method with geoscience foundation models for flood mapping. An Interpretable Ensemble Geoscience Foundation Model for flood mapping (IEGF) has been proposed. IEGF introduces an ensemble foundation model framework that achieves state-of-the-art performance on the flood mapping task while offering an interpretable decisionmaking process for human users. Together, these contributions advance the development of interpretable deep learning models for Earth observation and demonstrate that interpretability and high performance can be achieved simultaneously.